hustvl / ViG

ViG Model Card

Model Details

ViG is a generic backbone trained on the ImageNet-1K dataset for vision tasks.

• Developed by: HUST , Horizon Robotics
• Model type: A generic vision backbone based on the Gated Linear Attention (GLA) architecture.
• License: Non-commercial license

Model Sources

• Repository: https://github.com/hustvl/ViG
• Paper: https://arxiv.org/abs/2405.18425

Uses

The primary use of ViG is research on vision tasks, e.g., classification, segmentation, detection, and instance segmentation, with an GLA-based backbone. The primary intended users of the model are researchers and hobbyists in computer vision, machine learning, and artificial intelligence.

Training Details

ViG is pretrained on ImageNet-1K with classification supervision. The training data is around 1.3M images from ImageNet-1K dataset . See more details in this paper .

Evaluation

ViG is evaluated on ImageNet-1K val set, more details can be found in this paper .

Additional Information

Citation Information

 @article{vig,
  title={ViG: Linear-complexity Visual Sequence Learning with Gated Linear Attention},
  author={Bencheng Liao and Xinggang Wang and Lianghui Zhu and Qian Zhang and Chang Huang},
  journal={arXiv preprint arXiv:2405.18425},
  year={2024}
}